https://gcc.gnu.org/bugzilla/show_bug.cgi?id=61292
Andrew Pinski <pinskia at gcc dot gnu.org> changed: What |Removed |Added ---------------------------------------------------------------------------- Last reconfirmed| |2021-07-28 Status|UNCONFIRMED |NEW Summary|auto keyword to vector |auto keyword to reference |reference generates wrong |generates wrong alignment |alignment move (causing |move (causing runtime |runtime segfault) |segfault) Ever confirmed|0 |1 Keywords| |wrong-code --- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> --- This is unrelated to vector_size attribute even. Take: typedef float float32x4_t; typedef float __attribute__( ( aligned(1) ) ) float32x4a4_t; void add11(float32x4a4_t * x, float y, float32x4_t v) { auto & k1 = *(float32x4a4_t*)(x); auto & k2 = *(float32x4a4_t*)(x); k1 +=v; k2 += k1+v; } void add98(float32x4a4_t * x, float y, float32x4_t v) { float32x4a4_t & k1 = *(float32x4a4_t*)(x); float32x4a4_t & k2 = *(float32x4a4_t*)(x); k1 +=v; k2 += k1+v; } ------ CUT ----- With aarch64 at -O2 -mstrict-align, we get float loads/stores for add11 but correctly get the byte loads/stores for add98. Note I notice clang gets this wrong too.